Receiver-Conditioned Latent Communication gives 94% CacheBack
CacheBack filters multi-agent KV caches by receiver need, cutting 75% of state and raising FanOutQA accuracy 14.7 points.
The paper introduces receiver-conditioned latent communication so multi-agent systems transfer only the KV-cache state a receiving agent needs, instead of full caches that grow with context and agent count. CacheBack is a training-free method that uses the sender's attention weights to filter and compress that state. On FanOutQA with Qwen3, it removes 75% of the state agents would otherwise receive, improving accuracy by 14.7 percentage points and cutting median task-completion latency by 3.2x versus text communication. Comparable gains are reported across dense Transformers, Mamba-attention hybrids, and sliding-window attention models.
- Receiver-conditioned communication sends only KV-cache state the receiver needs.
- CacheBack is training-free and filters the sender cache using attention weights.
- On FanOutQA with Qwen3, it drops 75% of received state versus full cache transfer.
- Accuracy rose 14.7 points and median latency fell 3.2x versus text messages.
- Gains held on dense Transformers, Mamba-attention hybrids, and sliding-window models.
From the original article
1 min read at huggingface.co
Multi-agent systems distribute large contexts across agents that communicate to solve a task. Text messages are compact but require decoding and may omit evidence the receiving agent needs. Recent latent communication instead transfers KV caches.
By Maximillian Rossi, Prajwal Raghunath, Haoqing Xuan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.